HAL Id: hal-00813881
https://hal.inria.fr/hal-00813881
Submitted on 1 Jul 2013
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires
Tumor Growth Parameters Estimation and Source Localization From a Unique Time Point: Application to
Low-grade Gliomas
Islem Rekik, Stéphanie Allassonnière, Olivier Clatz, Ezequiel Geremia, Erin Stretton, Hervé Delingette, Nicholas Ayache
To cite this version:
Islem Rekik, Stéphanie Allassonnière, Olivier Clatz, Ezequiel Geremia, Erin Stretton, et al.. Tumor Growth Parameters Estimation and Source Localization From a Unique Time Point: Application to Low-grade Gliomas. Computer Vision and Image Understanding, Elsevier, 2013, 117 (3), pp.238–249.
�10.1016/j.cviu.2012.11.001�. �hal-00813881�
Tumor Growth Parameters Estimation and Source Localization From a Unique Time Point: Application to
Low-grade Gliomas
Islem Rekik
a, St´ ephanie Allassonni` ere
b, Olivier Clatz
a, Ezequiel Geremia
a, Erin Stretton
a, Herv´ e Delingette
a, Nicholas Ayache
aa
INRIA, Asclepios Project, 2004 Route des Lucioles, BP 93, F-06902 Sophia Antipolis Cedex, France
b
CMAP Ecole Polytechnique, Route de Saclay, 91128 Palaiseau France
Abstract
Coupling time series of MR Images with reaction-diffusion-based models has provided interesting ways to better understand the proliferative-invasive aspect of glial cells in tumors. In this paper, we address a different formula- tion of the inverse problem: from a single time point image of a non-swollen brain tumor, estimate the tumor source location and the diffusivity ratio between white and grey matter, while exploring the possibility to predict the further extent of the observed tumor at later time points in low-grade gliomas. The synthetic and clinical results show the stability of the located source and its varying distance from the tumor barycenter and how the esti- mated ratio controls the spikiness of the tumor.
Keywords: diffusivity ratio, source estimation, Eikonal Equation, reaction-diffusion glioma growth modeling
1. Introduction
Brain gliomas represent about 50% of all primary brain tumors [1] and can be classified according to their grade of malignancy. Low grade gliomas
Email addresses: islem.rekik@gmail.com (Islem Rekik),
stephanie.allassonniere@polytechnique.edu (St´ ephanie Allassonni` ere),
olivier.Clatz@inria.fr (Olivier Clatz), ezequiel.Geremia@inria.fr (Ezequiel
Geremia), erin.stretton@gmail.com (Erin Stretton), herve.Delingette@inria.fr
(Herv´ e Delingette), Nicholas.Ayache@inria.fr (Nicholas Ayache)
(LGG) are slow invaders of brain tissue as they keep growing for many years, presenting one of the most controversial decision treatment areas. High grade gliomas (HGG) remain unfortunately incurable with an average life expectancy of one year after its discovery, eventually creating symptoms due to an increase of the intracranial pressure or swelling around the tumor. The diagnosis of brain gliomas includes the analysis of various MRI sequences of the brain which partially reveal the tumor invasion. Based on those images and other clinical information, neurologists try to determine the grade of the gliomas and to estimate their current and further spatial extent and if possible their source location.
For more than a decade, mathematical models of brain tumors have been devised to help clinicians answer these questions. Microscopic models study the cellular mechanisms [2, 3, 4, 5] that explain the growth dynamics of gliomas at a microscopic scale. On the other hand, macroscopic models pioneered by Murray [6] describe the evolution of tumor cell density. How- ever, those quantities cannot be directly observed in clinical medical images, but it is commonly assumed that visible tumor boundaries correspond to an isovalue of this density. More realistic reaction-diffusion models have been proposed by Swanson et al. [7] based on the fact that tumor cells migrate faster on white matter fibre tracts myelin sheaths [8]. They have been fur- ther refined by Jbabdi et al. [9] and Clatz et al. [10] by considering an anisotropic diffusion in the white matter whose diffusion tensor is estimated from those acquired in DT-MRI.
A key issue for those models to answer clinical questions is their personal-
ization, i.e. the estimation of some patient-specific parameters from medical
images. The main parameters to be identified based on reaction-diffusion
models were pointed out in [11] as a combination of tumor diffusion tensors
in white and grey matter, its proliferation rate, its initial point and its initial
time. Several authors have estimated patient-specific parameters manually
[10] or through major model simplifications [12]. Colin et al. in [13] used a
reduced model based on Proper Orthogonal Decomposition (POD) in order to
identify growth parameters of pulmonary nodules in CT images. Konukoglu
et al. [14, 15, 16] have proposed an approach to automatically and accurately
personalize brain tumor models. They first remarked that given time series
of brain MR images, only the motion of a tumor front can be observed and
therefore only the three following parameters can be recovered : diffusivity in
the white d
wand grey matter d
gas well as the initial time T
0. Furthermore,
since the tumor cell density are only observed in MR images through visible
tumor boundaries corresponding to an isovalue, the reaction-diffusion equa- tions can be advantageously replaced by an Anisotropic Eikonal Equation (AEE) [16] which models the time at which a tumor front reaches a given point. By minimizing the distance between the segmented tumor and the simulated one, they were able to estimate uniquely those three parameters and test the prediction of future tumor evolution from at least a pair of images.
In this paper, we tackle a slightly different problem than the one ap- proached by previous authors. Instead of estimating the speed of tumor growth from a time series of images, we aim at characterizing the nature of the glioma, more precisely LGG, from a single MR image. Indeed, we hypoth- esized that the tumor shape is dependent on the proliferating or infiltrating nature of the tumor. Contrary to HGG where the presence of brain edema is common and usually associated with tumor malignancy, LGG are slowly growing tumors with a minimal surrounding edema [17, 18, 19]. Since our methodology main focus is on LGG, we will not consider the edema-induced mass effect in our further formulation of the tumor growth model. Therefore, the anatomical boundaries such as the ventricles’ will remain static as the tumor grows.
Given a segmented brain glioma from an MR image, we solve an inverse
problem in order to estimate the diffusivity ratio d
w/d
gand the tumor source
position. By localizing the tumor source and estimating the invaded tissue
characteristic using this ratio d
w/d
gour objective is to provide clinicians with
new indices that can be used for diagnosis from the first acquired MR images,
combined with a subsequent prediction of tumor invasive margins as it grows
from the initially observed boundary. This additional information may help
in surgical and/or radio-therapeutic treatment planning especially when it
comes to determining the margins for applying the therapy. The problem of
tumor seed localization was recently raised in [20], where a reaction-diffusion
based joint estimation of tumor evolution parameters was combined with a
multimodal deformable registration framework. This approach focuses on
the MR image registration with an atlas providing the estimation of the
initial seed location as a by-product. Later on, it was extended to a joint
segmentation and deformable registration between multi-sequence brain MR
image with gliomas and a probabilistic brain atlas of a healthy population [21,
22]. These recent publications targeting the localization of tumor seeds, the
quantification of the mass effect and its white matter diffusion coefficient was
based on a normal atlas.
In this paper, we investigate a different problem from registration or seg- mentation in low-grade gliomas. Our approach addresses a spatio-temporal tumor evolution with the estimation of the diffusivity ratio and the position of its source. Additionally, after solving the inverse problem, we evaluate the stability of the location of tumor sources over time and analyze the relation- ship between the tumor source and tumor barycenter as both locations have been assumed to match in past studies [23, 24, 25, 26, 27]. Finally, know- ing the diffusivity ratio d
w/d
gand the tumor source from a single image, we evaluate whether this information gives insights into predicting further tumor shape evolution in two distinct cases of low-grade gliomas.
2. Material and Method
2.1. MR Glioma images
In this work, we assume that for a given patient, one MR FLAIR image of the brain has been acquired, showing visible boundaries of glioma cells. We also assume that a Diffusion Tensor MR Image (DT-MRI) is also available and acquired at the same time as the FLAIR image. While FLAIR images are acquired in routine on patients with brain tumors, this is unfortunately not the case for DT-MRI. The extent of the tumor has been manually segmented in FLAIR images. Similarly, brain masks have been manually delineated on those images from which brain ventricles have been removed by a simple thresholding of the signal. Also white matter regions have been isolated by thresholding the voxels with the largest eigenvalue in the DT-MRI.
Objective: From the segmented tumor in the FLAIR image, our objective is to provide a quantitative analysis of the tumor shape which is not simply based on geometry (spheroid vs star-shaped) but based on simple biophysics growth principles. Indeed, two quantities are estimated in this analysis: the tumor source position and the diffusivity ratio between white matter and grey matter. This information may provide additional hints about the nature and the future progression of the tumor.
Data issues: Collecting LGG data with DT-MRIs is not a very straight-
forward task since diffusion MR is a quite recent technology [28], rarely used
in the common clinical practice and furthermore acquired DT-MRI may have
various anomalies like holes, low resolution and signal distortion. This is
particularly true around any tumor lesion. This lack of information may be
compensated by assuming the symmetry of the brain. For this, we perform
a symmetrization process to “reinitialize” the region where the tumor grew
and induced a diffusion signal distortion. Hence, for a proper simulation of tumor growth, we have corrected the tensor field by making the hypothesis that the DT MRI was originally symmetric with respect to the mid-sagittal plane. Thus the DT-MRI voxels corresponding to the largest extent of the tumor have been modified by symmetrizing and copying the voxels from the healthy brain hemisphere. However, this symmetrization process is prone to the following difficulties:
a- the DT MRI is asymmetric especially in white matter where about 50% of the contralateral tumor volume has zero Fractional Anisotropy (FA) values, while in the corresponding affected region where the tumor grew, the diffusion signal exists with a remarkable distortion (see Fig 1-a where the FA tumor map is darker than the contralateral part with a significant absence of symmetry).
b- even with a reliable DT-MRI symmetry, about 50% of the contralat- eral non-pathological symmetric volume to the tumor volume has no DT (i.e FA) signal. In Fig 2-a, the presence of large black hole in the FA signal in the contralateral part to the tumor invasion area presents a major barrier to diffusion tensor-guided glioma evolution simulation. To cope with this DT signal-missing problem, an interpolation algorithm based on isotropic diffu- sion (solving the heat equation) was applied to estimate the missing tensors from neighboring regions. This is done by applying a Gaussian convolution separately on the six components of the diffusion tensors.
Dataset: By excluding LGG cases with completely distorted or missing DTI, we succeeded to include four LGG patients. The first case, a.k.a pa- tient A, has developed a second grade astrocytoma classified as a low-grade glioma. Four successive time points of T2 flair MR images with a resolution of 0.99 × 1 × 2.16 mm
3were acquired and visible tumor boundaries were manually delineated by an expert. A DT-MRI image was also acquired at the first acquisition time point. As expected, the white matter fiber tracts are perturbed by the tumor growth and the DT-MRI signal near the tumor does not capture the original diffusion tensors of the brain at the onset of the disease. Therefore, we used the symmetrization process to reset the affected area to non-pathologically spatially deviated tensors (see Fig 3 where the white matter diffusion tracts were beautifully recovered).
The second case, a.k.a patient B, suffers from a low-grade glioma and
four MR images were acquired at distinct time points with a resolution of
0.89 × 0.97 × 0.97 mm
3. Only one DT-MRI was acquired during the initial
scanning process and the tumor region of the DT-MRI has been symmetrized
similarly to patient A. In addition, the DT-MRI includes small holes in the opposite region to the tumor that were successfully interpolated.
The third case, a.k.a patient C, with LGG has three acquisition time points with a resolution of 2 × 2 × 2 mm
3. In this particular case, displayed in Fig 1, we encountered difficulty (-a-) where the DT-MRI is not fully sym- metric.
The fourth case, a.k.a patient D, with a resolution of 2 × 2 × 2 mm
3was not symmetrized due to the large holes in the symmetric region to the tumor as show in Fig 2-a. Therefore, we interpolated the holes in the affected tissue without the symmetrization process.
2.2. Tumor Growth Modeling : from reaction-diffusion to Eikonal equations Glial cells dynamics are essentially governed by two biological phenom- ena : proliferation and invasion. They can be jointly modeled by a reaction- diffusion equation which describes the change over time of the normalized tumor cell density u:
∂u∂t
= ∇ . (D(x) ∇ u) + ρu(1 − u)
D ∇ u. n
∂Ω= 0 (1)
where ρ is the proliferation rate, D the local diffusion tensor, and n
∂Ωis the normal vector at the domain boundary surface ∂Ω. In the first equation, the proliferation of tumor cells follows a logistic growth parameterized by ρ whereas the tumor infiltration into neighboring neural fibers is captured by an anisotropic diffusion parameterized by D. The second equation indicates that there is no flux of tumor cells outside the domain Ω.
The diffusion tensor is a definite positive and symmetric 3 × 3 matrix whose value may be linked to Diffusion Tensor MRI (DT-MRI) [9]. Indeed, it characterizes the motility of tumor cells that is considered to be isotropic in grey matter but anisotropic in white matter. More precisely, the tumor diffusion tensor (TDT) may be written as D(x) = d
gI
3in grey matter, where d
gis the diffusivity coefficient.
In white matter, there are several approaches to link the TDT D(x)
with the DT-MRI signal D
water(x). Clatz et al. [10] proposed to have D(x)
proportional to D
water(x) whereas Jbabdi et al. [9] have introduced a formu-
lation which takes into account the possible equality of the two largest eigen-
values corresponding to a possible fiber crossing. Due to high anisotropies
of D
water(x) in most parts of the white matter, these two approaches lead
however to diffusivities that are much lower than d
gin the directions orthog- onal to the fibers, which is questionable. Furthermore, the high ratios of anisotropy encountered in those two expressions also lead to large computa- tional times.
In this paper, we propose to use the following white matter tumor diffu- sion tensor:
D(x) = V(x)[diag(e
1(x)d
w, d
g, d
g)]V(x)
T(2) where d
wis the white matter diffusivity coefficient, V(x) represents the ma- trix of sorted eigenvectors of D
water(x) and e
1(x), is the normalized largest eigenvalue (between 0 and 1) of D
water(x). With this choice, tensors have a non-homogeneous anisotropy ratio which is always less than d
w/d
gbut is maximized at the center of the white matter fibers and continuously decreases towards their boundaries. By simply dividing the duration of tumor evolution simulation of its propagating front using our adopted diffusion tensor 2 by the simulation duration as we used the diffusion tensor formulas presented in [9] (we precisely refer the reader to formulas A11 and A12), we have noticed that our choice speeds-up the computational time by a factor of 200 without any significant differences in performance. In fact, the use of a more nonlin- ear (more anisotropic) diffusion tensor field increases the computational time of the solution as the characteristic direction of the recursive anisotropic fast marching algorithm used to solve the AEE 5 becomes harder to find. This also may be explained by the fact that below a certain anisotropy ratio, the difference in tumor growth simulation is hardly noticeable.
The reaction-diffusion equation (1) is not practical when dealing with clin- ical images. Indeed, in MR images tumor cell density u cannot be observed but only the visible tumor boundary can. Hence, a front motion approxi- mation for the reaction-diffusion equation was introduced by Konukoglu et al. [16] assuming that the visible contour is associated with iso-density con- tour u = 0.4 [29]. They introduced an Anisotropic Eikonal Equation (AEE) describing the time T (x) at which the evolving tumor front passes through the location x. In its simplest form, the AEE writes as:
F √
∇ T
TD ∇ T = 1 with F = 2 √
ρ (3)
However, such approximation of equation 1 is too simple and Konukoglu et al. then proposed to account for the fact that the tumor front speed increases over time to reach an asymptotic value equal to 2
ρ n
TDn where
n is the normal direction of the front. The time-dependent speed of the
propagating tumor front was introduced in [16] as:
F = 4ρT − 3 2 √
ρT (4)
Furthermore, the front curvature κ
eff(x) also plays a role in the front speed as the front slows at high curvature points. This is especially important at the early stage of the tumor growth when the front is similar to a small sphere. With this additional hypothesis, the speed term becomes :
F = 4ρT − 3
2 √ ρT − 0.3 √ ρ
1 − e
−|κef f|/0.3√ρ(5) This last formulation is no longer a Hamilton-Jacobi equation due to the second-order curvature term and therefore cannot be solved by fast sweeping methods such as the Anisotropic Fast Marching (AFM) [14]. However, a multi-pass approach was proposed [16] to solve efficiently this equation by applying several times the AFM method while estimating the curvature front from previous iterations. The AFM method is recursive and the larger the tensor anisotropy the more iterations are needed to compute the character- istic direction of equation (5). Our white matter TDT of equation (2) limits the anisotropy ratio and therefore leads to reasonable computational times (typically few minutes for a tumor growth from a seed point).
2.3. Parameter estimation problem from a unique MR image
Based on the previously exposed mathematical model, we can simulate the growth of a glioma given its initial source S(x) for which we assume that T (S(x)) = 0. From this boundary condition and the knowledge of diffusivity d
w, d
gand proliferation rate ρ we can compute the time T (x) at which the visible tumor front reaches a given point. The isocontours of the field T (x) correspond to the successive shapes of the visible tumor boundary over time as shown in Fig 4. The speed on the front is not constant but its asymptotic value is 2
ρd
gin grey matter and 2 √
ρd
win white matter.
In this paper, we are interested in solving the following inverse problem:
given a visible tumor boundary S
Segin an MR image, can we extract the
growth parameters d
w, d
g, ρ, T
Obsand source location S(x) that best explain
the observed tumor boundary ? The duration T
Obsbetween the onset of the
tumor and the MR image acquisition is indeed also unknown.
Based on [16], it has been already established that several combinations of ρ, d
w, d
glead to the same front speed and therefore the same tumor growth simulations. Therefore, it is sufficient in this inverse problem to consider a fixed value of the proliferation rate ρ corresponding to the tumor grade and to estimate the remaining parameters. However, unlike [15, 16] this problem can be further simplified by realizing that the front speed cannot be estimated since T
Obsis unknown. If one multiplies the diffusivities by a scale factor α then one obtains the same isocontours for a propagation time divided by
√ α
1. This means that the simulated tumor isocontours do not depend on that absolute value of d
gand d
wbut on the diffusivity ratio:
r = d
wd
gIn the remainder, we will show that this ratio is related to the spikiness of the tumor.
A simple sensitivity analysis has led to conclusion that solving the in- verse tumor growth problem only depends on the following 2 parameters:
the source location S(x) and the spikiness index r. Here, the “spikiness index” represents a biology-driven estimated measure which quantifies the tortuousness of the boundary of the tumor as displayed on MR axial slices (see Figure 4-A), an index related to the frequency at which the tumor shape bends and twists. We consider that { S(x), r } appropriately characterize well a tumor extent if its visible boundary in MRI, S
Seg, is an isocontour of the simulated tumor growth initiated at S(x) with diffusivity ratios equal to r.
Therefore, we propose to estimate the patient specific parameters by mini- mizing the following criterion:
C
ρ(S(x, y, z), r) = 1 N T
x∈SSeg
(T (x) − T )
2(6) with
T = 1 N
x∈SSegT (x) (7)
1
This is not strictly true if one uses the speed term F taking into account the transient
speed as in Equation 4 or 5. However, the difference in simulations due to the absolute
value of the diffusivities was found to be negligeable.
where N is the number of points belonging to the manually delineated tu- mor boundary S
Seg. In these equations, T and C
ρare respectively the mean time value and time standard deviation computed over the tumor bound- ary. Our motivation to use this criterion to get good estimates of our un- knowns (S(x, y, z), r) derives from the fact that a tumor boundary (propa- gating front) is simultaneously defined as an isotime and an isosurface. Thus, to quantify how good is the estimation of the parameters guiding the spatio- temporal evolution of the tumor shape, we need to quantify how closely the simulated isosurface matches the observed one (manually delineated bound- ary). From a time perspective, this also implies that when measuring the time T at every point x of the manually delineated tumor isotime boundary, its value T (x) will be constant in the best case scenario where the simulated tumor front exactly matches the observed tumor boundary. Therefore, min- imizing the time standard variation criterion C
ρover the delineated tumor boundary S
Segwill guide the algorithm towards a better estimation of the two key parameters driving the invasive tumor front into fitting the succes- sive MR observed boundaries. Note that C
ρis normalized by T because the criterion should be made independent of the tumor front speed and therefore the mean time T .
In order to efficiently minimize the previously outlined criterion, we use the multidimensional unconstrained minimization algorithm without gradi- ent introduced by Powell in [30]. This algorithm suits our case since our parameters are bounded in both biological and geometrical spaces. More- over, the derivative of minimization criterion C
ρis not easy to compute. To better study the convergence of this algorithm and evaluate its outcome, 15 tests were performed using synthetic tumors. Further evaluation of this method was then studied using real data: two patients with LGG.
2.4. Synthetic Data: synthetic tumor generation process
In order to validate our parameter estimation, we produce synthetic tu-
mor MR images, where the initial tumor location and diffusivity ratio are
known. The procedure of generating synthetic tumors relies on choosing a
seed point S
0in either white or grey matter and the proliferation rate ρ to one
of the following values 0.008, 0.012, 0.024/day. The tumor front propagation
is simulated with a white matter diffusion rate d
wand grey matter diffusion
rate d
gwhose ratio varies between 1 and 100 (considered to be a biologically
valid range) using equation (5). The growth simulation is stopped at a speci-
fied time T
Obsthus leading to a time distance map as seen in Fig 4. Different
synthetic tumors were created at different anatomical locations with different sizes and asymptotic speeds of growth in both white and grey matter.
3. Results
3.1. Convexity of the minimization function: a convergence study
To check the convergence of the Powell minimization algorithm to the initially set parameters, we study the convexity of the minimization criterion C (see Equation (6)) for the four scalar parameters [S
x, S
y, S
z, r], writing separately the coordinates of the tumor source location. We proceed by alternatively fixing some parameters to their ground truth values and opti- mizing the remaining ones with the proposed minimization process. In Fig 5, we can clearly see the convexity of the minimization surface plotted after set- ting the diffusivity ratio r and one of the source coordinates to their right values.
The convexity of the minimization criterion C
ρwas successfully checked also when optimizing three parameters and setting the fourth one at its true value. The fact that the minimization criterion appears to be convex at the vicinity of the ground truth parameters is reassuring about the observability of the four parameters. However, the functional may still have local minima and practical optimization results will be discussed in the next sections.
3.2. Synthetic data
We evaluated our method on 15 synthetically generated tumors with a diffusivity ratio ranging from 1 to 100. The algorithm succeeded to identify the original tumor source with a mean error of 0.42mm and a standard devi- ation of 0.36mm. Moreover, it always converges to the real diffusivity ratio d
w/d
gwith a mean error equal to 0.18 and a standard deviation of 0.06.
Furthermore, we use this synthetic data to provide a better understanding of the diffusivity ratio r. The first row of Fig 6 shows how the shape of the tumor boundary can switch from “sphere-like”to “star-like”as the diffusivity ratio value jumps from 1 to 50. Although the displayed tumors in the third row are of different sizes, the correlation between tumors’ shape spikiness and the value of the diffusivity ratio r ∈ [1 − 100] is clear.
Besides, we also notice from the second row of Fig 6 similar tumor con-
tours in terms of spikiness and irregularity when fixing the diffusivity ratio
at the same value and varying the location of the initial tumor seed. This
confirms that the diffusivity ratio value r controls the spikiness of tumor shape.
3.3. Clinical data
3.3.1. Evaluation criteria
In clinical data, no ground truth values of the parameters are available and therefore additional criteria must be introduced in order to assess the quality of the parameter estimation.
After the optimization of criterion 6, we extract the simulated isocon- tour which is closest from the visible tumor boundary in MRI. Computing symmetric distances between the two surfaces provides a quantitative infor- mation about how well the tumor shape can be explained by the proposed tumor growth model. We detail below the proposed approach.
We extract the closest isocontour defined by time T
Obsˆ by optimizing C
isoT ime(50%)(T ) :
T
Obsˆ = min
T
C
isoT ime(50%)(T ) (8)
The criterion C
isoT ime(50%)(T ) is defined as the median symmetric distance between the visible tumour boundary S
Segand the isosurface ˆ S
isoT imeat time T . More precisely, for each voxel of ˆ S
isoT ime(resp. S
Seg) its closest distance from S
Seg(resp. ˆ S
isoT ime) is computed through a distance map and added to a list. The median value of that list is then taken as C
isoT ime(50%)(T ). The optimization of the functional is done with the Powell algorithm [30] already used for solving the general inverse problem and requiring few estimations of the functional. The range of time value for the optimization is constrained to be in the range [T − δ; T − δ] where δ is the standard deviation:
δ =
x∈SSeg
(T (x) − T )
2N
We also compute other robust distance criteria C
isoT ime(y%)(T ) by taking the y% quantile of the symmetric distances.
Once the closest isochrone surface ˆ S
isoT imeis estimated, we define the
symmetric distance between (S, ˆ S
isoT ime) as the first evaluation criterion:
C
symDist(S
Seg, S ˆ
isoT ime) =
x∈SSeg
min d(x, S ˆ
isoT ime) +
x∈SˆisoT ime